Tingyu Wang

Comments on 03/22: - [x] explain / add a link to "linearly independent" - should discuss the choice between `reshape`+`@` and `dot` for mat-vec - correct the usage of "system...

> I believe we don't want to add new stuff to dgl.contrib, any suggestions on the directory to use? @jermainewang > I will recommend putting the example in a folder...

> The achieved accuracy is different from DGL. Why is that? Just saw this comment. I think it is because `num_bases` is set to `num_rels` in the current DGL example....

> Thanks for the reminder. I will break it into two PRs.

> As we discussed earlier during the meeting, please update accordingly. I also suggest splitting this PR into two, one for NN module and one for example. It is generally...

> Make sure you update the PyG version in `dependencies.yaml` too. Updated now, though `cugraph_pyg_dev` not being used under the current setup.

> Sorry I messed this PR up. I'll approve once the unintentional changes I made are reverted. @alexbarghi-nv The diff has been fixed, feel free to take a look.

> @tingyu66 we need changes in the cuGraph-ops integration for PyG 2.5. > > Right now the following tests are failing: > > ``` > FAILED python/cugraph-pyg/cugraph_pyg/tests/nn/test_gat_conv.py::test_gat_conv_equality[basic_pyg_graph_1-True-8-1-True-False-True] - AttributeError: 'NoneType'...

We have two modes in the layer: 1). users pass in concatenated `edge_emb` that is going to be consumed by MLP or 2). users pass in a_scalars, b_scalars, edge_scalars separately...

Just to clarify, is the ask here to better document the caveat when using mode 2, right? With mode 1, edge_emb could be constructed in arbitrary ways, and I don't...